At one of the PET workshops {these discussed much of the academic background to the technology behind TOR, Mixminion etc} someone presented a paper on how long a user would wait for a reply to a web request. Up jumps an attendee who says he and some colleagues had already done a paper on this, and the answer was 5 seconds. Ooops.
Now for a 5 second maximum response time it is technologically infeasable to implement an untrusted onion network which resists attack by a global persistent threat like the NSA or GCHQ.The APT just temporally correlates data exiting the endpoints.
The network can introduce timing jitter and packet size standardisation or variation in order to make this harder, but with cost-limited dummy traffic and a maximum 5 seconds response time it can't reliably stop it.
It cain't be done.
This is such an old discussion that I am also not up to date on, but I don't see why snail mail and constant bandwidth were not effective counterarguments to user expectations and timing correlation.